Run ThinLTO release smoke tests in parallel

[helly25]

Default opt tests and ThinLTO release smoke builds now run in parallel instead of serially in the same Linux and macOS jobs. This preserves validation of the exact shipped binaries while aiming to remove the release-build tail from required CI wall time.

AG;DR

Change

• Keep the complete ordinary test suite in the existing default matrix cells.
• Move the tagged ThinLTO release tests, lean/full release builds, staging, executable smoke checks, and PR artifact uploads into independent Linux and macOS release smoke jobs.
• Give release builds their own Bazel disk-cache namespace so incompatible opt and ThinLTO outputs do not contend.
• Make the aggregate done job require the new release-smoke matrix.

Timing hypothesis

The final PR 777 run serialized these phases:

Platform	Default tests	ThinLTO release phase	Serialized total
Linux	14m32s	6m41s	21m13s
macOS	16m53s	4m08s	21m01s

The split is worthwhile if each new release-smoke cell finishes no later than its matching default-test cell, after accounting for independent setup and cache behavior. This PR's CI run is the acceptance measurement. If that condition does not hold, the next experiment is running the complete suite under ThinLTO and comparing its wall time.

Validation

• tools/release_prep_test.sh
• git diff --check origin/main...HEAD
• pre-commit run --files .github/workflows/main.yml

No test scope or release transformation is removed; only scheduling and cache ownership change.

[helly25]

Superseded by #780 after timing showed that one complete production ThinLTO test build has comparable wall time, substantially lower runner cost, and avoids configuration drift.